Your dense retriever scores well offline and badly on real user questions over a technical corpus, and you have no relevance labels. What is the bug and what is the cheapest fix?

The asymmetry between document-document training and question-document inference is the root cause. You can diagnose this by comparing mean cosines of document-document pairs versus question-document pairs. HyDE is the ideal solution here because it is label-free and can be implemented by generating a single hypothetical answer to serve as a query vector, effectively aligning the distribution.